Covariate Adjustment: Prespecify Prognostic Variables Before Unblinding

For a randomized, parallel-group drug or biologic trial, consider adjusting the primary analysis for a small, justified set of prognostic baseline variables. Identify them from prior evidence, account for randomization strata, and prespecify the complete analysis before unblinding comparative data. An unadjusted primary analysis remains acceptable under FDA guidance; adjustment is an opportunity to improve precision, not a universal regulatory requirement. FDA final guidance, sections II–III.

Imagine a protocol team preparing a two-arm Phase 3 superiority trial. The proposed endpoint is a continuous follow-up score, and randomization will be stratified by baseline severity. The immediate decision is whether the baseline score and severity belong in the primary model, how each enters it, and what happens if a baseline measurement is absent. Answering only “use analysis of covariance (ANCOVA)” leaves important choices unresolved.

Use the protocol to establish the principal analysis choices and the statistical analysis plan (SAP) to make the implementation executable. FDA emphasizes the prespecified primary analysis when reviewing a trial. EMA also allows justified updates before unblinding when prior knowledge changes; prespecification does not mean that a plan can never be amended. Document the reason, timing, and information available for any revision. FDA section III.A; EMA section 6.1.

FDA lists Adjusting for Covariates in Randomized Clinical Trials for Drugs and Biological Products as a Final Level 1 guidance issued in May 2023, docket FDA-2019-D-0934. It covers randomized parallel-group superiority and noninferiority trials of drugs and biological products. These are nonbinding recommendations. The announcement confirms that the final text replaced the May 21, 2021 draft and clarified standard errors, stratification, interactions, and additional methods. FDA guidance page; FDA final-guidance announcement.

The efficiency rationale starts with prognosis. Randomization balances baseline characteristics in expectation, while adjustment can explain outcome variability that remains between participants. Covariates with little prognostic value may give no precision gain or may reduce precision. Adding several strongly correlated variables can offer less additional benefit than adding variables that contribute different prognostic information. FDA sections II and III.A.

• The guidance does not address covariate use to control confounding in nonrandomized trials. Its recommendations should not be transferred to observational comparisons as if randomization were present.

• It does not address covariates in models accounting for missing outcome data, or covariate adjustment for longitudinal repeated-measures data. The repeated-measures exclusion is broader than models containing time-varying covariates.

• Bayesian and machine-learning (ML) methods for covariate adjustment are outside its scope. An exclusion from this guidance is not a declaration that a method is prohibited.

FDA recommends review-division discussion for complex covariate-adaptive randomization, data-adaptive covariate selection, adjustment within an adaptive design, many covariates relative to sample size, or a factor with many levels such as study site. It also identifies percentage-change outcomes and noninferiority comparisons using ratios of group means for discussion. Keep these flags visible in the rationale sheet rather than treating a generic guidance citation as sufficient justification. FDA section III.A.

Begin with disease knowledge, published literature, prior studies such as Phase 2 trials, and suitable prognostic indices. For each candidate, explain why it should predict this endpoint in this trial population. Include the baseline value of a continuous outcome when available and ordinarily account for non-administrative stratification factors. A variable’s convenience in the database is not itself a prognostic rationale. FDA section III.A; EMA sections 5.1–5.6.

EMA says that retrospective selection of covariates most strongly associated with the current trial’s outcome should be avoided in confirmatory trials. Observing baseline imbalance after the fact is also not an appropriate reason to add a variable to the primary model. A sensitivity analysis may examine such an imbalance while preserving the planned primary analysis. FDA likewise emphasizes the prespecified primary analysis over post-hoc models with different covariates. EMA sections 5.4 and 6.1; FDA section III.A.

There is no agency rule limiting every trial to two, three, or four covariates. EMA explicitly gives no formal maximum. Consider the number of parameters as well as the number of named variables: a many-level factor or flexible curve can consume several parameters. The aim is a defensible model for the available sample, not compliance with an invented numerical quota. EMA section 6.2.

Shao and colleagues’ 2026 manuscript, available as arXiv version 3, reports an empirical comparison across 50 public randomized-trial datasets, 29,094 participants, and 574 treatment–outcome comparisons. Median variance reductions were 13.3% for continuous outcomes and 4.6% for binary outcomes. Parsimonious regression performed as well as or better than more complex methods in the evaluated settings; machine-learning estimators added no precision overall and had more computational failures for binary outcomes. These are results from a retrospective public-data benchmark, not guaranteed gains for a new trial. The binary analyses targeted marginal risk differences, and the evidence should not be generalized to every endpoint or estimator. Shao et al., abstract and methods.

For planning, justify the assumed prognostic strength in the intended population. FDA permits sample-size calculations based on adjusted or unadjusted methods, noting that the latter will often be more conservative. Do not automatically reduce recruitment by a median percentage from another collection of trials. The effect size, outcome variation, missingness, allocation, and planned estimator still need a trial-specific basis. FDA section III.A.

A covariate-adjustment model should generally include the variables used to stratify randomization. EMA makes an exception for factors used solely for administrative purposes and asks sponsors to explain mismatches between stratification and analysis. If errors create “actual” and “as-randomized” strata, FDA allows either definition when prespecified. The guidance does not designate one as the universal default. FDA section III.A; EMA section 5.2.

Sullivan and colleagues reviewed 105 trials, including 32 that stratified on an underlying continuous variable. None reported adjustment for its underlying continuous values in the primary analysis. This does not mean all 32 adjusted for categories: ten used an unadjusted overall comparison, and three had unclear adjustment. Their simulations favored appropriately modelling continuous values, with flexible forms helping guard against misspecification. Categorization could lose information; misspecification was especially consequential in settings with missing outcomes. These findings support a prespecified functional-form decision rather than an automatic rule that every covariate must enter linearly. Sullivan et al., sections 2–6.

For a variable categorized for randomization, record both the allocation categories and the planned analysis representation. Justify retaining the continuous values and any nonlinear terms; explain how the analysis reflects the restricted allocation. EMA’s older guideline favors simple forms when prior knowledge is limited, whereas the methods paper emphasizes flexible modelling. State that difference openly and support the chosen form instead of presenting one modelling convention as universally agreed. EMA section 6.3; Sullivan et al.

In linear models, FDA recommends considering Huber–White sandwich standard errors when treatment-by-covariate interactions are absent; an appropriate nonparametric bootstrap is another option. The text also accepts nominal standard errors in two-arm trials with 1:1 randomization even under model misspecification. It separately recommends that standard-error computation account for stratified randomization. Selecting a software option labelled “robust” does not by itself demonstrate that both requirements have been addressed. FDA section III.B.

Specify the variance formula or bootstrap procedure, how it respects the allocation scheme, and how finite-sample performance will be assessed. The Shao manuscript cautions that sandwich estimates may be inaccurate in small samples; it does not establish a universal threshold of 50 participants. Neither regulator document establishes one variance method that automatically satisfies every multiregional design. Shao et al., practical recommendations; FDA section III.B.

For continuous outcomes in the randomized setting described by FDA, ordinary least squares adjustment can estimate the unconditional average treatment effect even when the working linear model is misspecified. Appropriate inference still matters. This robustness is not a general solution for missing outcomes or nonrandomized comparisons. EMA notes that with baseline included in a standard linear model, analysis of the raw follow-up outcome and analysis of change from baseline give identical treatment-effect estimates. FDA section III.B; EMA section 5.6.

Nonlinear models need an additional decision. Odds ratios and hazard ratios can be non-collapsible: a conditional comparison within covariate-defined groups can differ from the marginal population comparison even without confounding. FDA’s hypothetical example has a conditional odds ratio of 8.0 in each of two strata and a marginal odds ratio of 4.8. Those are illustrative calculations, not observed trial results or individual patient predictions. FDA section III.C and Table 1.

• A conditional effect describes a treatment comparison within levels of specified baseline covariates, under the relevant model assumptions. It is not a prediction of an individual person’s causal benefit.

• An unconditional or marginal effect compares outcomes under treatment and control averaged over the target population. Choosing it requires an estimator and summary scale aligned to that question.

ICH E9(R1) frames the treatment-effect question through the estimand, including its population-level summary. It does not require every analysis to target a marginal effect, nor does covariate adjustment replace the other estimand attributes. FDA asks sponsors to prespecify whether the target is conditional or unconditional and to discuss primary nonlinear-regression proposals for conditional effects, including model-assumption assessments, with the review division. ICH E9(R1), sections A.3 and A.6; FDA section III.C.

For a marginal binary-outcome comparison, FDA illustrates standardized g-computation: fit a logistic working model, predict each participant’s response probability under each treatment condition, average the predictions separately, then contrast the two average risks on the chosen scale. It also describes inverse probability of treatment weighting as a reliable approach to unconditional effects. These are examples, not an exhaustive menu; the variance calculation and assumptions must match the estimator. The binary recipe should not be copied into a survival analysis without a suitable survival method. FDA section III.C.

The table below summarizes the May 2023 FDA text and the 2015 EMA guideline available as of October 23, 2026. The two texts differ in date and detail; this comparison is an editorial interpretation of their published recommendations, not a harmonized agency position.

Sources for the comparison: FDA sections III.A–C and EMA sections 4.2, 5.4, 6.1, 6.4–6.5, and 7.3. A simpler common model may be useful, but “write to the stricter intersection” is not proof that every design issue has been resolved. In particular, many small centres, nonlinear model assumptions, and allocation-dependent inference need their own justification. FDA final guidance; EMA guideline.

A missing baseline covariate and a missing outcome are different problems. FDA’s explicit scope exclusion concerns covariates in models for missing outcomes; it should not be paraphrased as a ban on all missing-data methods. Its text does not provide a missing-baseline algorithm. EMA section 6.5 says that excluding patients because their covariate information is missing violates the intention-to-treat (ITT) principle and is not generally acceptable, so handling should be considered at planning. FDA section I; EMA section 6.5.

Chang and colleagues’ 2023 methods study supports imputation of missing baseline values together with missingness indicators and the required indicator-by-covariate interactions in the adjustment or propensity model. It studies continuous and binary outcomes, with outcomes complete or missing under specified assumptions; survival settings need further research. This is a method to evaluate for the proposed trial, not a universal regulatory default. Its findings do not establish that any imputation scheme plus one indicator guarantees identical estimates, correct inference, or full ITT compliance. Chang et al.

In the SAP, specify how the imputation constant is calculated, which indicators and interactions enter the model, and how the procedure links to the chosen estimator and variance method. An indicator-by-covariate interaction is distinct from a treatment-by-covariate interaction; assess the implemented model rather than assuming all interaction terms have the same regulatory implications. Retaining a participant after a missing baseline measurement does not recover a missing endpoint. Endpoint missingness and intercurrent events still need estimand-aligned plans.

The following sheet is a proposed working template. Its example is synthetic: a two-arm randomized trial with a continuous follow-up score and severity-stratified allocation. No disease-specific prognostic effect, prior-trial correlation, sample size, or validated model performance is being asserted. Replace each placeholder with identifiable trial evidence before sign-off.

Before sign-off, have the statistician verify that every proposed primary-model term has a rationale and that the written model matches the executable specification. Record supporting analyses and any prespecified response to failed model assumptions. ICH E9 cautions against adjusting the main analysis for variables measured after randomization that may be affected by treatment; a baseline rationale sheet should not quietly become a collection of post-treatment predictors. ICH E9, section 5.7.

For the broader division of protocol and SAP responsibilities, use the site’s SAP endpoint-definition guide. For the full treatment-effect question and missing-endpoint assumptions, see the estimand guide and missing-data sensitivity-analysis guide. The specific decision here remains which baseline variables belong in the primary model and how their implementation is justified before unblinding.

• FDA Guidance: Adjusting for Covariates in Randomized Clinical Trials for Drugs and Biological Products

• FDA Guidance for Industry: Adjusting for Covariates in Randomized Clinical Trials for Drugs and Biological Products (Final May 2023)

• FDA Announcement: Final Guidance on Adjusting for Covariates in Randomized Clinical Trials

• EMA Guideline on Adjustment for Baseline Covariates in Clinical Trials (EMA/CHMP/295050/2013)

• ICH E9(R1) Addendum on Estimands and Sensitivity Analysis in Clinical Trials

• ICH E9 Statistical Principles for Clinical Trials

• Chang et al. (2023) Covariate Adjustment in Randomized Clinical Trials with Missing Covariate and Outcome Data

• Sullivan et al. (2024) Categorisation of Continuous Covariates for Stratified Randomisation: How Should We Adjust?

• Shao et al. (2026) How Should Covariates Be Handled in Randomized Trials? Empirical Evidence from 50 Trials